Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Delete hstack and hslice patterns in pattern matching logic #37

Merged
merged 1 commit into from
Oct 8, 2024
Merged

Delete hstack and hslice patterns in pattern matching logic #37

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 0 additions & 50 deletions python/tvm/relay/op/contrib/forge/forge.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,22 +59,6 @@ def dense_to_matmul():
weight_t = is_op('transpose')(weight)
return is_op('nn.dense')(data, weight_t)

def reshape_transpose_to_hslice():
act = wildcard()
act_r = is_op('reshape')(act)
return is_op('transpose')(act_r)

def transpose_reshape_to_hstack():
act = wildcard()
act_t = is_op("transpose")(act)
return is_op("reshape")(act_t)

def transpose_reshape_reshape_to_hstack():
act = wildcard()
act_t = is_op("transpose")(act)
rshp = is_op("reshape")(act_t)
return is_op("reshape")(rshp)

def reshape_to_vstack():
act = wildcard()
return is_op('reshape')(act)
Expand Down Expand Up @@ -147,11 +131,6 @@ def channel_last_resize():
@register_pattern_table("forge")
def pattern_table():
matmul = ("forge.matmul", dense_to_matmul())
hslice = ("forge.hslice", reshape_transpose_to_hslice(), is_reshape_transpose_hslice)
hstack = [
("forge.hstack", transpose_reshape_to_hstack(), is_transpose_reshape_hstack),
("forge.hstack", transpose_reshape_reshape_to_hstack(), is_transpose_reshape_reshape_hstack)
]
binary_stack = [
("forge.binary_stack", stack_reshape_reshape_to_binary_stack(), is_stack_reshape_reshape_to_binary_stack),
("forge.binary_stack", concat_reshape_reshape_to_binary_stack(), is_concat_reshape_reshape_to_binary_stack),
Expand All @@ -165,9 +144,7 @@ def pattern_table():
# channel_last_maxpool2d = ("forge.channel_last_maxpool", channel_last_maxpool())
channel_last_resize2d = ("forge.channel_last_resize2d", channel_last_resize())
forge_patterns = [
*hstack,
*binary_stack,
hslice,
adv_index,
matmul,
forge_conv2d_with_bias,
Expand Down Expand Up @@ -220,8 +197,6 @@ def pattern_table():
# Forge
"forge.adv_index",
"forge.binary_stack",
"forge.hslice",
"forge.hstack",
]

# TM CPU Fallback ops which should not be included in fallback
Expand Down Expand Up @@ -361,31 +336,6 @@ def visit_call(self, call):

logger.trace("CPU fallback on linear")
return super().visit(tvm.relay.nn.dense(arg0, arg1))

if "Composite" in call.op.attrs and call.op.attrs["Composite"] == "forge_cpudevice.hslice":
# Get hslice input
if isinstance(call.args[0], tvm.relay.expr.Call):
arg0 = call.args[0]
else:
return super().visit_call(call)

# Extract reshape and transpose attributes
if isinstance(call.op.body, tvm.relay.expr.Call) and call.op.body.op.name == "transpose" and \
isinstance(call.op.body.args[0], tvm.relay.expr.Call) and call.op.body.args[0].op.name == "reshape":
transpose = call.op.body
transpose_axes = transpose.attrs.axes
reshape = call.op.body.args[0]
reshape_new_shape = reshape.attrs.newshape
else:
return super().visit_call(call)

# Unwrap hslice function into composite ops
ops = tvm.relay.reshape(arg0, reshape_new_shape)
ops = tvm.relay.transpose(ops, transpose_axes)

logger.trace("CPU fallback on hslice")
return ops

if "Composite" in call.op.attrs and call.op.attrs["Composite"] == "forge_cpudevice.binary_stack" and "PartitionedFromPattern" in call.op.attrs and call.op.attrs["PartitionedFromPattern"] == "Tuple_stack_reshape_":
# Get binary_stack input
arg0 = call.args[0] if len(call.args) > 0 else None
Expand Down